Researchers trained an OpenAI rival in half an hour for less than $50
The researchers managed to create a low-cost AI reasoning model rivaling OpenAI’s in just 26 minutes, as outlined in a paper published last week. The model, called s1, was refined using a small dataset of 1,000 questions and for under $50, according to TechCrunch.
Using Distillation to Train the Model
To achieve this, researchers at Stanford and the University of Washington utilized a method known as distillation. This technique allows smaller models to learn from the answers generated by larger models. In this case, they refined the s1 model using answers from Google’s AI reasoning model, Gemini 2.0 Flash Thinking Experimental.
The researchers based s1 on Qwen2.5, an open-source model from Alibaba Cloud. They trained the model on a dataset of 1,000 questions, which they found to be more effective than a larger dataset of 59,000 questions. The training was done on just 16 Nvidia H100 GPUs.
Test-Time Scaling Technique
The s1 model also incorporates a technique called test-time scaling, allowing the model to deliberate for a longer duration before producing an answer. By adding "Wait" to the model's response, researchers encouraged the model to double-check its reasoning steps, leading to more accurate answers.
Implications for the Industry
The emergence of smaller and more affordable AI models poses a threat to major players in the industry. Companies like OpenAI, Microsoft, Meta, and Google may no longer need to invest billions of dollars in training AI models or building massive data centers filled with Nvidia GPUs.
OpenAI's o1 reasoning model follows a similar approach, and other startups like DeepSeek have attempted to replicate this with their own models. However, OpenAI has accused DeepSeek of violating its terms of service by allegedly distilling information from its models to create a competitor.
According to the researchers, s1 exceeds o1-preview on competition math questions by up to 27%, showcasing its efficiency and accuracy in reasoning tasks.
As the landscape of AI continues to evolve, the development of cost-effective and high-performing models like s1 could revolutionize the industry and drive innovation in AI technology.
